Design And Implementation Of An Auxiliary Detection System For Colorectal Lesions Based On Deep Learning

Being one of the most prevalent malignant tumors in China,colorectal cancer is a serious threat to people’s health and its incidence has been on the rise in recent years.Therefore,early screening and diagnosis play an important role in saving patients’ lives.Colonoscopy combined with whole abdominal CT scan is a widely used method for early screening of high-risk groups of colorectal cancer and early diagnosis of primary colorectal cancer.On the one hand,colorectal endoscopy can observe the lesions inside the intestine;on the other hand,CT can help determine the location of lesions and the grading of cancer.Pathologists often undertake a huge workload when viewing pathological images,and due to the varying medical conditions and the different standards of pathologists across the country,there can be missed or misdiagnosed cases.In this context,in order to help doctors diagnose colorectal cancer lesions efficiently,reduce the work pressure of pathologists,and thus reduce the rate of missed diagnosis and misdiagnosis of colorectal cancer,this paper selects colorectal endoscopic images and whole abdominal CT plain scan images as the research objects,conducts research on target detection and target segmentation algorithms based on deep learning,and applies the algorithm modules of these two parts to an online medical assistant diagnosis system with certain data security.The main work of this paper includes the following aspects:(1)Research on the algorithm of colorectal endoscopic polyp detection.Colorectal endoscopic polyp images have the characteristics of small targets and inconspicuous features,and high requirements for real-time image processing.Based on Yolov5 s,this paper proposes an improved model,Polyp＿yolov5s,which is specially designed for polyp targets.The recognition accuracy is effectively improved by adding prediction heads,feature enhancement,and optimizing prediction frames.The experimental results show that the improved model is much better than the yolov5 s model before the improvement in recognition accuracy,and has better accuracy and faster processing speed than the yolov5 l with a deeper network.(2)Research on the segmentation algorithm of colorectal mass in plain CT scan of the whole abdominal cavity.This paper fine-tunes the nn Unet model under the Pytorch deep learning framework to complete image segmentation of abdominal CT colorectal masses,and also validates the segmentation effect of nn Unet on liver organs and liver masses based on the characteristics of primary colorectal cancer which is prone to liver metastasis.(3)Based on the two algorithms mentioned above,this paper designs and implements an online colorectal lesion auxiliary detection system with certain concurrency and data security according to the actual needs.The system has functions such as image import and display,doctor user management,patient information management,authority control,abdominal CT image colorectal mass segmentation,endoscopic image target rapid detection,APM server status monitoring and log recording,and construction of display reports.The online system can be deployed in the remote cloud or the hospital intranet,which may improve the lack of medical resources in areas where the level of hardware is insufficient,in line with China’s "Internet +" medical development trend.After testing the system,various functions realized are normal and can provide auxiliary testing for doctors,which has some practical application value.